JP3435714B2 - Neural network type additional learning device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network type additional learning device, and is particularly effective when used as a control device capable of learning individual preferences in an air conditioner.

[0002]

2. Description of the Related Art Conventionally, there has been a control device for an air conditioner or the like that realizes control suitable for an individual's preference by learning the individual's preference and usage. For example, JP-A-3-
In 177742, the automatic control function of the air conditioner is
It has been disclosed that a control function operated by a user is provided with a function of learning by a neural network, and air conditioning control suitable for the user's preference is realized. Here, the neural network operates from 0 to 1 from the minimum value to the maximum value of the input signal.
Learning is performed on the dimensionless value.

[0003]

By the way, in a vehicle air conditioner, the range of input signals such as the outside temperature and the inside temperature is wide, and if learning is performed on the input signals in the entire range, the learning input / output relationship is changed. There is a problem that learning is difficult to converge due to the large number of teacher data shown, and a problem that learning time is long until the number of times of learning is large, so that the learning performance by the neural network is deteriorated.

Since the range of the input signal when the vehicle occupant operates it to reset the air-conditioning state is relatively narrower than the range of all the input signals, the present invention uses the range of the input signal and the neural network. By distinguishing from the input range to be learned,
It is an object of the present invention to provide a neural network type additional learning device capable of preventing the deterioration of learning performance.

[0005]

In order to achieve the above object, the present invention calculates a target outlet temperature as shown in FIG.
And control the air flow to reach this target outlet temperature.
A neural network type additional learning device for adjustment, which is a signal input means 11 for inputting a signal, and the signal input means 11
The input signal input from the
Determination means for determining whether the signal within a predetermined range Te 1
5 and a first control pattern which is input when the input signal is determined by the determination means 15 to be within the predetermined range, and which includes the input signal and an output signal output in response to the input signal. Is learned by a neural network, and the first control unit 121 that outputs an output signal toward the controlled object based on the first control pattern, and the determining unit 15 determine that the input signal is not within the predetermined range. A second control unit 122 that outputs an output signal toward the controlled object based on a second control pattern that is input when the input signal is output and that is output according to the input signal. An output changing unit 13 for changing the output signals output by the first control unit 121 and the second control unit 122 into an arbitrary output signal, and the output changing unit 13. When the output signal is changed, the storage unit 14 stores the additional learning pattern including the output signal after the change and the input signal when the output is changed, and the first control unit 121 includes the storage unit 14. Neural network type addition for re-learning the first control pattern based on the additional learning pattern stored in 14 corresponding to the input signal within the predetermined range and the first control pattern corresponding to the input signal within the predetermined range It employs a learning device.

[0006]

According to the neural network type additional learning device of the present invention having the above-mentioned structure, when the judging means 15 judges that the input signal is within the predetermined range, the first control section 121 makes the first When the determination unit 15 determines that the input signal is out of the predetermined range, the second control unit 122 outputs it based on the second control pattern.

When the output signal is changed by the output changing means 13, the changed output signal and the input signal when the output is changed are stored in the storage unit 14 as additional learning patterns. First
The control unit 121 of the first control pattern of the first control pattern is stored in the storage unit 14 based on the additional learning pattern corresponding to the input signal within the predetermined range and the first control pattern corresponding to the input signal within the predetermined range. Do additional learning.

[0008]

EXAMPLE An example in which the neural network type additional learning device of the present invention is used in a vehicle air conditioner will be described below.

FIG. 2 shows an example of a vehicle air conditioning control device according to the present invention. This air conditioning control device has an air duct 40 equipped in the vehicle. Inside the air duct 40, from the upstream side to the downstream side, an inside / outside air switching damper 41, a blower 42, an evaporator 43,
An air mix damper 44, a heater core 45, and an outlet switching damper 46 are provided. The inside / outside air switching damper 41 is switched to the outside air introduction position (the position shown by the solid line in FIG. 2) by the servomotor 41a to introduce outside air into the air duct 40 from the outside air introduction port 47, while the inside air introduction position ( The inside air in the passenger compartment of the vehicle is introduced into the air duct 40 through the inside air introduction port 48 by switching to the position shown by the dashed line in FIG.

The blower 42 is connected to the outside air inlet 4 according to the rotation speed of the blower motor M driven by the drive circuit 42a.
The outside air from 7 or the inside air from the inside air introduction port 48 is introduced as an air flow through the inside / outside air switching damper 41 and is blown to the evaporator 43. The evaporator 43 cools the air flow from the blower 42 according to the operation of the refrigeration cycle of the air conditioning control device. The air mix damper 44 is driven by the servo motor 44a and causes the cooling air flow from the evaporator 43 to flow into the heater core 45 according to the opening degree. The remaining cooling air flow bypasses the heater core 45 and flows toward the outlet switching damper 46. The heater core 45 heats the inflowing airflow by exchanging heat with the inflowing cooling airflow in accordance with the temperature of the cooling water from the engine cooling system of the vehicle, and causes it to flow toward the outlet switching damper 46. On the downstream side of the heater core 45, the airflow that has passed through the heater core 45 and is heated and the airflow that bypassed the heater core 45 are air-mixed.

The blower outlet switching damper 46 is switched to the first switching position (the position shown by the solid line in FIG. 2) during the ventilation mode (VENT) of the air conditioning controller under the drive of the servo motor 46a. ,
An air flow is blown from the VENT outlet 13 of the air duct 40 toward the center of the passenger compartment. Further, the air outlet switching damper 46 is switched to the second switching position (the position indicated by the alternate long and short dash line in FIG. 2) during the heat mode (HEAT) of the air conditioning control position under the drive of the servo motor 46a. 40 FOOT outlets 1
An air flow is blown from 4 toward the lower part of the passenger compartment. Also,
The outlet switching damper 46 is a servo motor 46a.
Driven by the air conditioning control position in the bi-level mode (B / L), it is switched to the third switching position (the position indicated by the chain double-dashed line in FIG. 2) and the air outlets 13 and 1 are connected.
An air flow is blown from 4 toward the center of the vehicle interior and downward.

The sensor group 30 corresponding to the signal input means of the present invention comprises an inside air temperature sensor 61, an outside air temperature sensor 62, a solar radiation sensor 63, an outlet temperature sensor 64, and a water temperature sensor 65. The inside air temperature sensor 61 detects the actual temperature inside the passenger compartment of the vehicle and generates it as an inside air temperature detection signal. The outside air temperature sensor 62 detects the actual temperature of the outside air of the vehicle and generates it as an outside air temperature detection signal. The solar radiation sensor 63 detects the actual amount of solar radiation incident on the vehicle interior and generates it as a solar radiation detection signal. The outlet temperature sensor 64 generates the actual temperature at the outlet of the evaporator 43 as an outlet temperature detection signal. The water temperature sensor 65 detects the actual temperature of the cooling water of the engine cooling system of the vehicle and generates it as a water temperature detection signal.

The A / D converter 66 has an inside air temperature detection signal from the inside air temperature sensor 61, an outside air temperature detection signal from the outside air temperature sensor 62, a solar radiation detection signal from the solar radiation sensor 63, and an outlet temperature from the outlet temperature sensor 64. detection signal, respectively water temperature detection signal from the water temperature sensor 65 to digital conversion, the inside temperature T _R,
It is generated as a digital signal representing the outside air temperature T _AM , the amount of solar radiation T _S , the outlet temperature Te, and the water temperature Tw.

The microcomputer 67 operates the drive circuit 42 according to the flow charts shown in FIGS.
a, the arithmetic processing necessary for controlling the drive of each servo motor 41a, 44a, 46a is executed.

However, the above computer program is stored in the ROM of the microcomputer 67 in advance. The microcomputer 67 is powered by the battery B via the ignition switch iG of the vehicle and starts operating.

The microcomputer 67 has a learnable control section 121 corresponding to the first control section of the present invention.
And a fixed control unit 122 corresponding to the second control unit of the present invention.
And the storage unit 14 are connected to each other, and the learnable control unit 1
21 and the fixed control unit 122, the air conditioner is controlled, and the teacher data (corresponding to the additional learning pattern of the present invention) indicating the input / output relationship to be additionally learned is stored in the storage unit 14.

The learnable control section 121 is composed of the neural network 80 shown in FIG. 8, and is based on the teacher data (corresponding to the first control pattern of the present invention) showing the input / output relationship of the control characteristics shown in FIG. The control characteristics are already learned, and a signal is input when the target outlet temperature TAO, which will be described in detail later, is in a predetermined range A to B (hereinafter, referred to as a learning region) to control the blower air volume. Is.

The fixed control unit 122 is a learning enable control unit 12.
It is configured by the neural network 80 in the same manner as in No. 1, and in the control characteristic shown in FIG.
When O is less than A or greater than B (hereinafter referred to as a non-learning area), the teacher data indicating the input / output relationship of the control characteristics within this range (the second control pattern of the present invention). (Hereinafter, referred to as a characteristic outside the learning area) is stored.

As shown in FIG. 8, the neural network 80 used in the learning control section 121 has information as to whether the output signal output when a certain input signal is given is a correct signal or an erroneous signal ( The input layer 8 provided in the neural network 80 by inputting the teacher signal).
Weights of the connection 86 between the neurons 85 in the first, second and third intermediate layers 82 and 83 and the output layer 84 (sypnus weight)
It is equipped with an error backpropagation learning (backpropagation) function that corrects, and it is possible to increase the probability of outputting the correct answer when a new signal is input by repeatedly "learning". is there.

The weight of the connection 86 between the neurons 85 is
This represents the strength of the connection between neurons, and at the connection part where the neurons 85 are connected by the connection 86 having a large weight, the signal from the neuron 85 in the layer on the front side as viewed from the input side has a large amplitude. It becomes a signal, and the signal flows to the neuron 85 on the downstream side. In addition, in the connection portion with a small weight, the signal from the neuron 85 becomes a signal with a small amplitude, and the signal flows to the neuron 85 on the downstream side.

The operation panel P is provided on the instrument panel in front of the driver's seat in the passenger compartment, and as shown in FIG. 3, each part such as the air mix damper 44 and the outlet switching damper 46 has a predetermined control pattern. Auto switch 5 to set the auto mode for automatic control
1. An OFF switch 52 for inputting an operation stop command of the air conditioning control device, that is, for stopping the drive circuit 42a that drives the blower 42, and switching between operation and stop of the refrigeration cycle, that is, operation of a compressor (not shown) An air conditioner switch 53 for switching the stop, an inside / outside air switching switch 54 for switching between inside air circulation and outside air introduction, a temperature setting switch 55 for setting a target temperature in the vehicle interior, and a vehicle interior set by the temperature setting switch 55. 56 for displaying the target temperature of the blower, an air volume setting switch 57 for setting the amount of air blown from the blower 42 in any of four stages from weak to strong, an outlet setting switch 58 for setting the outlet, etc. Is equipped with. An indicator 57d is provided above the air volume setting switch 57 to indicate which air volume is being set.

The temperature setting switch 55, the air volume setting switch 57, the outlet setting switch 58 and the like provided on the panel P correspond to the output changing means of the present invention. Further, the microcomputer 67 includes the judging means of the present invention.

Next, the microcomputer 67 shown in FIG.
The basic control by the will be described with reference to FIG. The microcomputer 67 starts the control in step 300 when the ignition switch iG is turned on, proceeds to step 310, and sets various conversions and initial values such as flags.

At the next step 320, the inside air temperature sensor 6
1, the sensor signals from the outside air temperature sensor 62, the solar radiation sensor 63, and the like, and the states of the operation switches on the panel P, such as the temperature setting switch 55, are input, and the process proceeds to step 330.

In step 330, the target outlet temperature (TAO) of the air blown into the vehicle compartment is calculated according to the following equation 1 based on the environmental conditions input in step 320.

[0026]

[Number 1] _{TAO = K SET × T SET -K} R × T R -K AM ×
_{T AM -K S × T S +} C However _{K SET, K R, K AM} , K S is a coefficient, C is a constant, T _SET is the set temperature, T _R is the inside air temperature, T _AM is the outside air temperature, T _S Is the amount of solar radiation.

Next, in step 340, the opening degree of the air mix damper 44 with respect to TAO is calculated, and an actuator (not shown) is controlled via the servo motor 44a so as to reach this opening degree. Controls the temperature of the air sent from 14 toward the passenger compartment.

Next, in step 350, the air flow rate is calculated, and the fan 42 fixed to the blower motor M is rotated through the drive circuit 42a to control the air flow rate. Next, the routine proceeds to step 360, where the inside / outside air introduction ratio by the inside / outside air switching damper 41 is calculated, and an actuator (not shown) is controlled via the servo motor 41a.

Next, in step 370, the state of the outlet mode is calculated, and the outlet switching damper 46 is controlled via the servo motor 46a. Next, in step 380, ON / OFF control of a compressor (not shown) is performed. When the control of this compressor ends, step 3
Returning to 20, the above process is repeated. Therefore, repeat T
AO has been calculated and updated.

Next, the embodiment of the present invention will be described in detail with reference to FIGS. 5 and 7 by taking the blower air flow rate control in step 350 as an example. When the air mix control in step 340 is completed, the blower control process shown in the detailed flowchart of FIG. 5 is performed.

In the blower air volume control, first, step 350
In step 1, it is judged by the OFF switch 52 of the panel P whether or not the control of the air conditioner is OFF, and if it is OFF, the routine proceeds to step 3502.

At step 3502, OFF processing such as stopping the blower motor M and completely erasing the display 57d showing the blower control state is performed, and the routine proceeds to step 360. on the other hand,
If it is determined in step 3501 that the air conditioner control is ON, the process proceeds to step 3503. Step 350
In step 3, whether or not the blower is automatically controlled is judged from the setting state of the AUTO switch 51 of the panel P, and if it is set to automatic control, the process proceeds to step 3504.

At step 3504, the target outlet temperature TA
It is determined whether or not O is within the learning area of the predetermined ranges A to B shown in FIG. 7, and if it is within the learning area, the learnable control unit 1
The blower voltage is determined according to the control characteristic 21 (the characteristic shown in FIG. 7A), and the blower drive circuit 42a is controlled. Further, when it is in the non-learning area outside the predetermined range A to B, the blower voltage is determined according to the control characteristic of the fixed control unit 122 (the characteristic shown in FIG. 7B), and the blower drive circuit 42a is controlled. Then, the process proceeds to step 360.

On the other hand, if it is determined in step 3503 that the blower is manually controlled, the process proceeds to step 3505. In step 3503, the setting state of the air volume setting switch 57 of the panel P is judged. For example, if the air volume M1 setting switch is ON, a blower control signal for realizing the air volume M1 is output to the blower drive circuit 42a and displayed. Manual processing is performed to turn on the indicator indicating the air volume M1 from the device 57d, and the process proceeds to step 3506.

In step 3506, it is determined whether or not the current manual process is the first process that has transitioned from the automatic control mode. If it is not the first process, the process proceeds to step 360 as "no". On the other hand, in the first process, that is, in the automatic control, it is determined that the occupant determines that the air volume does not suit his / her preference, and the air volume is set manually to the desired air volume using the air volume setting switch 57 of the panel P, and then the first air flow comes in. If you judge that, "yes
s ”and the processing proceeds to step 3507.

In step 3507, it is determined whether or not the environmental condition from the sensor group 30 during the manual setting is within the learning region, and the blower characteristic is determined by the additional control pattern consisting of the environmental condition and the changed blower air volume. (For example, the characteristic (a) of FIG. 7) is learned and changed (for example, the characteristic (c) of FIG. 7) and the process proceeds to step 360.

The learning process in step 3507 will be described in detail with reference to the flowchart of FIG. 6 and the control characteristics of FIG. First, in step 3520, the step at the time when the manual setting by the air volume setting switch 57 is performed.
Target air temperature TAO _n calculated in step 330 and preset values A and B indicating the range of the preset learning region
Are compared with each other to determine whether or not they are within the learning area.
The target outlet temperature TAO _n during manual setting is outside the learning range (TA
O _n <A, or TAO _n> B) and it is determined the process proceeds to step 3521. On the other hand, when the target outlet temperature TAO _n is within the learning region, the determination is “YES” and step 352
Go to 2.

In step 3521, the values of the learning area set values A and B set in advance are set to TAO at the time of manual setting.
Update the learning area by changing the value to include _n . For example, in the control characteristic of FIG. 7, when the TAO _n value at the time of manual setting is C larger than B, the learning area setting value B is set to C
To include B '.

Next, at step 3522, an additional control pattern consisting of the environmental condition from the sensor group 30 at the time of manual setting by the air volume setting switch 57 and the blower air volume after the change is stored in the storage section 14, and the routine proceeds to step 3523.

In step 3523, the additional control pattern data stored in the storage unit 14 is standardized to 0 to 1 and used as learning data for the neural network. Here, the standardization of learning data will be described. The value of the target outlet temperature TAO by the above formula has a range of about ± 400 ° C. depending on the combination of the set temperature Tset, the inside air temperature Tr, the outside air temperature Tam, and the solar radiation amount Ts. On the other hand, the inflection points a, b, c and d of the blower control characteristic shown in FIG.
It is within the range of 0 ° C. Outside this range, the blower drive circuit 4
2a is configured to be uniformly driven with the maximum voltage (maximum air volume). In other words, the region that is likely to be changed according to human preference is within the range of about -50 to 100 ° C, and outside this region is the extreme environmental condition,
It is an area where human preference is unlikely to occur.

Therefore, the learning data (additional control pattern) of the neural network is limited to the learning data (TAO between A and B) in the learning area including the inflection points a, b, c and d of the blower control characteristic. As a result, the control characteristics in the area can be relearned based on the learning data in this area and the teacher data indicating the input / output relationship of the control characteristics in the learning area shown in FIG. It becomes easier and the learning time to converge can be shortened. It is possible to achieve both simplification of the structure such as reduction of the number of intermediate layers of the neural network and improvement of learning ability such as improvement of learning resolution.

The learning data standardized in the above manner is used as teacher data (data indicating the input / output relationship) of the neural network in step 3524. Step 35
In 24, the connection weight between the neurons of the neural network is adjusted by a known back propagation method so that the output is as close as possible to the teacher data, and as a result, a learning characteristic such as the characteristic (c) shown in FIG. 7 is obtained. .

This preference characteristic is reflected in the blower automatic control step 3504 after the above processing, and thereafter, the preference automatic control reflecting the changed value of the output during the manual operation is carried out.

In the above embodiment, the application example of the blower control is shown, but the present invention can be applied to other control such as outlet control. Further, in the above-mentioned embodiment, an example in which the learning method is applied to the learning of the neural network by the back propagation method is shown, but a neural network by another learning side may be used.

Further, in the above-mentioned embodiment, the intermediate layer is a neural network having a two-layer structure, but the number of intermediate layers is not particularly limited and one layer may be used. In the above embodiment, it is determined in step 3520 whether or not the target outlet temperature TAOn at the time of changing the blower air volume is within the control area, and if it is outside the control area, the control area is set in step 3521. Although it was configured to repair, without resetting the control area,
The control areas of the learnable control 121 and the fixed control unit 122 may be fixed. In this case, the storage unit 3
The additional control pattern stored in 4 is the TA in the control area.
Only the pattern corresponding to On may be stored and learning may be performed according to this pattern.

In the above embodiment, the fixed controller 12
2 is composed of a neural network, the fixed controller 12
Since No. 2 does not perform additional learning, it may be configured by an electronic circuit including a microcomputer.

[0047]

As described above, in the neural network type additional learning device of the present invention, the control unit for controlling the output is
It is composed of a first control unit that can perform additional learning and controls based on the first control pattern, and a second control unit that controls based on the second control pattern. When the input signal is within a predetermined range, the first control unit The control section controls, and when the input signal is outside the predetermined range, the second control section controls.

The first control section comprises an output signal changed by the output changing means and an input signal at the time of changing the output.
Additional learning is performed based on the additional learning pattern corresponding to the input signal within the predetermined range.

Therefore, as compared with learning for the entire range of the input signal, only the additional learning pattern corresponding to the input signal within the predetermined range and the first control pattern corresponding to the input signal within the predetermined range. Because it re-learns based on
The first control pattern and the additional learning pattern learned by the first control unit are reduced, and the learning is easily converged.
The learning time can be shortened, and thus the learning performance can be improved.

[Brief description of drawings]

FIG. 1 is a diagram corresponding to the claims of the present invention.

FIG. 2 shows an overall configuration diagram when the neural network type additional learning device of the present invention is used in a vehicle air conditioner.

FIG. 3 is a view showing a panel P.

FIG. 4 is a flowchart showing control of a vehicle air conditioner using the neural network type additional learning device of the present invention.

FIG. 5 is a flowchart showing control of a vehicle air conditioner using the neural network type additional learning device of the present invention.

FIG. 6 is a flowchart showing control of a vehicle air conditioner using the neural network-type additional learning device of the present invention.

FIG. 7 is a diagram showing a control pattern using the neural network-type additional learning device of the present invention.

FIG. 8 is a diagram showing a neural network used in the present invention.

[Explanation of symbols]

11 Signal input means 12 Control unit 121 First control unit (learnable control unit) 122 Second control unit (fixed control unit) 13 Output change means 14 Memory 15 Judgment means 80 Neural Net

Front page continuation (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06N 1/00-7/08 G06G 7/60 F24F 11/00-11/08 G05B 13/00-13/04 JST file (JOIS) CSDB (Japan Patent Office)

Claims (2)

(57) [Claims] 1. A target outlet temperature is calculated to obtain the target outlet temperature.
Neural network for air conditioning that controls the air flow
A Tsu preparative additional learning apparatus, a signal input means for inputting a signal, the input signal is the target blowing inputted from the signal input means
Determining means for determining whether the signal is within a predetermined range based on the output temperature; and inputting when the input signal is determined to be a signal within the predetermined range by the determining means, the input signal A first control unit configured to learn a first control pattern composed of an output signal output according to the input signal by a neural network, and output the output signal toward a controlled object based on the first control pattern, The control based on a second control pattern that is input when the determination means determines that the input signal is not within a predetermined range, and that includes the input signal and an output signal that is output according to the input signal. A second control unit that outputs an output signal toward the target; and an output changing unit that changes the output signals output by the first control unit and the second control unit to arbitrary output signals A storage unit that stores an additional learning pattern including the output signal after the change and the input signal when the output is changed, when the output signal is changed by the output changing unit. Re-learns the first control pattern based on the additional learning pattern stored in the storage unit and corresponding to the input signal within the predetermined range and the first control pattern corresponding to the input signal within the predetermined range. Neural network type additional learning device. 2. The input signal at the time of changing the output signal is included when the input signal at the time of changing the output signal by the output changing means is judged not to be within the predetermined range by the judging means. The neural network type additional learning device according to claim 1, further comprising update means for updating the predetermined range.